Awesome Speaker Diarization

https://github.com/wq2012/awesome-diarization

Admin

Awesome Speaker Diarization is a curated list of papers, libraries, datasets, and other resources...

Tokens:13,086

Snippets:245

Trust Score:9.5

License:Apache-2.0

Update:2 months ago

Show doc for...

Tokens:

Raw

### End-to-End Model Hyperparameters

Source: https://github.com/wq2012/awesome-diarization/blob/master/_autodocs/quick-reference.md

General configuration parameters for end-to-end diarization architectures.

```text
encoder: RNN, Transformer, or Conformer
output_dim: 256-512 (embedding size)
num_speakers: 2-10 (max in training)
training_loss: permutation-invariant or attractor-based
```

--------------------------------

### Spectral Clustering Hyperparameters

Source: https://github.com/wq2012/awesome-diarization/blob/master/_autodocs/quick-reference.md

Configuration parameters for spectral clustering algorithms in diarization tasks.

```text
affinity_matrix_type: 'cosine' or 'precomputed'
n_clusters: auto-tuned or 2-10 (typical range)
p_neighbors: 10-20 (for refinement)
threshold: 0.5-0.8 (for threshold stopping criterion)
```

--------------------------------

### UIS-RNN Hyperparameters

Source: https://github.com/wq2012/awesome-diarization/blob/master/_autodocs/quick-reference.md

Standard hyperparameter ranges for training UIS-RNN models.

```text
hidden_size: 512 or 1024
num_layers: 1-3
learning_rate: 0.001-0.0001
batch_size: 32-64
```

--------------------------------

### PLDA Hyperparameters

Source: https://github.com/wq2012/awesome-diarization/blob/master/_autodocs/quick-reference.md

Configuration settings for Probabilistic Linear Discriminant Analysis (PLDA).

```text
n_components: 200-400 (latent dimension)
n_classes: varies by training set
Nit: 10-50 (EM iterations)
```

--------------------------------

### Diarization Error Rate (DER) Formula

Source: https://github.com/wq2012/awesome-diarization/blob/master/_autodocs/quick-reference.md

Mathematical definition of the primary diarization evaluation metric.

```text
DER = (Ms + Fs + Cs) / Ref_duration × 100%

Ms = Speaker miss (false negative)
Fs = False alarm (speaker wrongly detected)
Cs = Speaker confusion (label mismatch)
Ref_duration = Total reference duration in seconds
```

Tokens:

Raw